When benchmarking AH it is very important to pick a particular terrain, a particular location, a particular view, and even a particular zoom level. The fps can vary wildly with those parameters (at least for those of us not blessed with high speed computers that max out the fps all the time). My examples:
Mmesa terrain at A1, Shift F1, details max/max, offline, 4xAA/8xAF, 1280x1024, 512 textures.
T34 in hanger, gunsight view zoomed max*: 9600XT = 42, 7600GT = 80
T34 driver view, head out the window, unzoomed: 9600XT = 29, 7600GT = 28 (more trees to render so CPU bogs down)
T34 driver view, head out, zoomed max: 9600XT = 53, 7600GT = 58 (somewhat fewer trees, CPU speeds up)
Doing the same with SHIFT F4 gives more trees and lower fps: 35/73, 23/21, 46/48.
Note that under many view conditions I am limited not by my graphics card, but by my CPU (P4 Northwood 2.8GHz w 1.5GB RAM), and it was a mistake to upgrade my graphics card (but then it was only $100).
*this used to bring AH to it's knees until HTC made a significant improvement in the rendering algorithm, that's why I include it.